Hugging Face BlogT1·Models
62 pts
Ai2 has released XYBench, a benchmark of over 8,000 questions designed to test whether AI assistants can recognize the real goal behind a misguided question, the classic XY problem, instead of answering the surface request. The dataset and code are open-sourced, with current coverage limited to English-language and mainly procedural problems.